obs-ffmpeg: Add automatic CUDA texture sharing fallback

obsproject · Jan 31, 2024 · 646fc7b · 646fc7b
1 parent 3a15ace
commit 646fc7b
Show file tree

Hide file tree

Showing 3 changed files with 51 additions and 19 deletions.
diff --git a/plugins/obs-ffmpeg/obs-nvenc-helpers.c b/plugins/obs-ffmpeg/obs-nvenc-helpers.c
@@ -277,6 +277,7 @@ static const cuda_function cuda_functions[] = {
 	{offsetof(CudaFunctions, cuMemcpy2D), "cuMemcpy2D_v2"},
 
 #ifndef _WIN32
+	{offsetof(CudaFunctions, cuGLGetDevices), "cuGLGetDevices_v2"},
 	{offsetof(CudaFunctions, cuGraphicsGLRegisterImage),
 	 "cuGraphicsGLRegisterImage"},
 	{offsetof(CudaFunctions, cuGraphicsUnregisterResource),

diff --git a/plugins/obs-ffmpeg/obs-nvenc.c b/plugins/obs-ffmpeg/obs-nvenc.c
@@ -28,9 +28,6 @@
 /* we cannot guarantee structures haven't changed, so purposely break on
  * version change to force the programmer to update or remove backward
  * compatibility NVENC code. */
-#if CONFIGURED_NVENC_VER != NVENCAPI_VERSION
-#error NVENC version changed, update or remove NVENC compatibility code
-#endif
 
 #undef NVENCAPI_STRUCT_VERSION
 #define NVENCAPI_STRUCT_VERSION(ver)                            \
@@ -1229,19 +1226,63 @@ static bool init_cuda_surfaces(struct nvenc_data *enc)
 	return true;
 }
 
-static bool init_cuda_ctx(struct nvenc_data *enc, obs_data_t *settings)
+static bool init_cuda_ctx(struct nvenc_data *enc, obs_data_t *settings,
+			  const bool texture)
 {
-	int count;
+	int count, gpu;
 	CUdevice device;
-	const int gpu = (int)obs_data_get_int(settings, "gpu");
+	bool cuda_override;
+
+	/* Allow CUDA device override for texture encoders (experimental) */
+	if (obs_data_has_user_value(settings, "cuda_device")) {
+		gpu = (int)obs_data_get_int(settings, "cuda_device");
+		cuda_override = true;
+	} else {
+		gpu = (int)obs_data_get_int(settings, "gpu");
+		cuda_override = false;
+	}
 
 	CU_FAILED(cu->cuInit(0))
 	CU_FAILED(cu->cuDeviceGetCount(&count))
 	if (!count) {
 		NV_FAIL("No CUDA devices found");
 		return false;
 	}
+#ifdef _WIN32
 	CU_FAILED(cu->cuDeviceGet(&device, gpu))
+#else
+	if (!texture || cuda_override) {
+		CU_FAILED(cu->cuDeviceGet(&device, gpu))
+	} else {
+		unsigned int ctx_count = 0;
+		CUdevice devices[2];
+
+		obs_enter_graphics();
+		CUresult res = cu->cuGLGetDevices(&ctx_count, devices, 2,
+						  CU_GL_DEVICE_LIST_ALL);
+		obs_leave_graphics();
+
+		if (res != CUDA_SUCCESS || !ctx_count) {
+			/* CUDA_ERROR_INVALID_GRAPHICS_CONTEXT should be treated
+			 * as non-fatal fallback (probably running on iGPU). */
+			if (res == 219) {
+				info("Not running on NVIDIA GPU, falling back to non-texture encoder");
+			} else {
+				error("Failed to get a CUDA device for the current OpenGL context: %d",
+				      res);
+			}
+			return false;
+		}
+
+		/* Documentation indicates this should only ever happen with SLI, i.e. never for OBS. */
+		if (ctx_count > 1) {
+			warn("Got more than one CUDA devices for OpenGL context, this is untested.");
+		}
+
+		device = devices[0];
+		debug("Loading up CUDA on device %u", device);
+	}
+#endif
 	CU_FAILED(cu->cuCtxCreate(&enc->cu_ctx, 0, device))
 	CU_FAILED(cu->cuCtxPopCurrent(NULL))
 
@@ -1388,7 +1429,7 @@ static void *nvenc_create_internal(enum codec_type codec, obs_data_t *settings,
 #ifdef _WIN32
 		!texture &&
 #endif
-		!init_cuda_ctx(enc, settings)) {
+		!init_cuda_ctx(enc, settings, texture)) {
 		goto fail;
 	}
 	if (get_nvenc_ver() == COMPATIBILITY_VERSION) {
@@ -2292,12 +2333,8 @@ struct obs_encoder_info h264_nvenc_soft_info = {
 	.id = "h264_fallback_nvenc",
 	.codec = "h264",
 	.type = OBS_ENCODER_VIDEO,
-#ifdef _WIN32
 	.caps = OBS_ENCODER_CAP_DYN_BITRATE | OBS_ENCODER_CAP_ROI |
 		OBS_ENCODER_CAP_INTERNAL,
-#else
-	.caps = OBS_ENCODER_CAP_DYN_BITRATE | OBS_ENCODER_CAP_ROI,
-#endif
 	.get_name = h264_nvenc_soft_get_name,
 	.create = h264_nvenc_soft_create,
 	.destroy = nvenc_destroy,
@@ -2315,12 +2352,8 @@ struct obs_encoder_info hevc_nvenc_soft_info = {
 	.id = "hevc_fallback_nvenc",
 	.codec = "hevc",
 	.type = OBS_ENCODER_VIDEO,
-#ifdef _WIN32
 	.caps = OBS_ENCODER_CAP_DYN_BITRATE | OBS_ENCODER_CAP_ROI |
 		OBS_ENCODER_CAP_INTERNAL,
-#else
-	.caps = OBS_ENCODER_CAP_DYN_BITRATE | OBS_ENCODER_CAP_ROI,
-#endif
 	.get_name = hevc_nvenc_soft_get_name,
 	.create = hevc_nvenc_soft_create,
 	.destroy = nvenc_destroy,
@@ -2338,12 +2371,8 @@ struct obs_encoder_info av1_nvenc_soft_info = {
 	.id = "av1_fallback_nvenc",
 	.codec = "av1",
 	.type = OBS_ENCODER_VIDEO,
-#ifdef _WIN32
 	.caps = OBS_ENCODER_CAP_DYN_BITRATE | OBS_ENCODER_CAP_ROI |
 		OBS_ENCODER_CAP_INTERNAL,
-#else
-	.caps = OBS_ENCODER_CAP_DYN_BITRATE | OBS_ENCODER_CAP_ROI,
-#endif
 	.get_name = av1_nvenc_soft_get_name,
 	.create = av1_nvenc_soft_create,
 	.destroy = nvenc_destroy,

diff --git a/plugins/obs-ffmpeg/obs-nvenc.h b/plugins/obs-ffmpeg/obs-nvenc.h
@@ -31,6 +31,7 @@ typedef struct CudaFunctions {
 	tcuMemcpy2D_v2 *cuMemcpy2D;
 
 #ifndef _WIN32
+	tcuGLGetDevices_v2 *cuGLGetDevices;
 	tcuGraphicsGLRegisterImage *cuGraphicsGLRegisterImage;
 	tcuGraphicsUnregisterResource *cuGraphicsUnregisterResource;
 	tcuGraphicsMapResources *cuGraphicsMapResources;
@@ -47,6 +48,7 @@ extern const char *nv_error_name(NVENCSTATUS err);
 extern NV_ENCODE_API_FUNCTION_LIST nv;
 extern NV_CREATE_INSTANCE_FUNC nv_create_instance;
 extern CudaFunctions *cu;
+extern int cuTexDeviceIdx;
 extern uint32_t get_nvenc_ver(void);
 extern bool init_nvenc(obs_encoder_t *encoder);
 extern bool init_cuda(obs_encoder_t *encoder);